Agentic Engineering Framework

Purpose & goals

The Agentic Engineering Framework is a governance and continuity harness wrapped around AI coding agents — Claude Code, Cursor, Copilot and the like. It exists because a capable actor operating in a shared environment needs the same things whether it is human or machine: clear direction, awareness of its context, awareness of the resources it is spending, awareness of the impact of its changes, and engagement with the work. Those are the same structural requirements observed across twenty-five years of enterprise IT governance — now applied to AI agents instead of human operators. The framework coordinates the agent; it does not do the agent’s work for it.

Modules & architecture

The design is organised as five requirement “layers,” each backed by a concrete mechanism rather than a good intention:

• A task gate. A pre-action hook refuses to modify a file when no task is active — nothing gets done without a task, and that rule is enforced structurally, not by discipline.
• Three memory layers. Working memory holds the live session; project memory accumulates patterns, decisions, learnings, and known gaps; episodic memory condenses each finished task into a history the next session can read.
• A context-budget gate. It reads the live session transcript and hands work over cleanly before the agent runs out of room, so long runs end in a checkpoint rather than a cliff.
• Impact analysis. Blast-radius and component-fabric tooling map what a change touches before it ships.
• Audit & coordination. A compliance audit and cross-agent coordination keep multiple actors honest.

A single command-line entry point, fw, routes to a set of focused agents — task-create, handover, context, healing, resume, audit, git, fabric.

Functionality & technology

In practice the framework enforces task traceability, structural gates, session continuity, and audit trails, using git-hook-style enforcement, semantic recall over accumulated learnings, and value-and-impact scoring to decide what is worth doing. It is written primarily in Shell, integrates the Model Context Protocol, and is deliberately provider-neutral — the vendor-agnostic guide lives in FRAMEWORK.md, with the Claude Code integration in CLAUDE.md. It is candid about itself: the agent that wrote its README tripped its own task gate in the process.